Long-running agents
Maintain reasoning and tool state across multi-step workflows where a model must plan, act, inspect results, and continue.
Independent tool overview
Trinity-Large-Thinking at a glance
Trinity-Large-Thinking is Arcee AI's open-weight 398B sparse MoE reasoning model for tool calling, multi-step planning, coding, and long-running agent loops.
Visit the official Trinity-Large-Thinking site ↗
- Developer
- Arcee AI
- Architecture
- 398B sparse MoE; about 13B active
- Expert routing
- 4 of 256 experts plus 1 shared
- Model context
- Up to 512K tokens
- OpenRouter context
- 262,144 tokens
- Core capabilities
- Reasoning, tool calling, and coding
- License
- OpenMDW-1.1
- Languages listed
- 11
Overview
What Trinity-Large-Thinking is
Trinity-Large-Thinking is the reasoning-optimized checkpoint in Arcee AI's Trinity Large family. It has roughly 398 billion total parameters but activates about 13 billion per token through 4-of-256 expert routing, aiming to combine frontier-scale capacity with more efficient inference than a similarly sized dense model.
The model is post-trained with extended chain-of-thought and agentic reinforcement learning for multi-step planning, tool calling, coding, and customer-service-style workflows. It emits explicit reasoning in think blocks or a reasoning_content field, and Arcee warns that this reasoning must be preserved in message history for reliable multi-turn agent behavior.
Arcee's model card describes a 512K extended context window, while the current OpenRouter endpoint exposes 262,144 tokens. The downloadable weights use the permissive OpenMDW-1.1 license, but a 398B checkpoint remains operationally demanding even with quantized variants, so hosted API access is the practical starting point for most teams.
Use cases
Who Trinity-Large-Thinking is best for
The strongest fit depends on the job you need the product to complete, not the size of its feature list.
Tool-heavy customer operations
Orchestrate structured lookups and actions across service, telecom, travel, and other process-driven workflows.
Agentic coding
Power coding agents that need planning, command execution, repository work, and iterative verification.
Open-weight deployment
Run or customize a frontier-scale reasoning model when the organization needs control over weights and infrastructure.
Cost-sensitive reasoning APIs
Evaluate a lower-priced hosted alternative to closed frontier reasoning models for high token volumes.
Capabilities
Core Trinity-Large-Thinking features
Sparse 398B mixture of experts
Routes each token through four experts plus one shared expert, activating about 13B parameters rather than the full model.
Explicit reasoning traces
Produces extended thinking that can be returned separately from the final answer and preserved across turns.
Native tool calling
Supports structured tool calls for agent loops, with documented parsers and OpenAI-compatible serving patterns.
Long context
The checkpoint is extended to 512K tokens, though hosted providers may expose a smaller window.
Agent-framework compatibility
Arcee documents compatibility with OpenClaw and Hermes Agent, and recommends vLLM for agentic deployment.
Multiple weight formats
Official releases include the main checkpoint plus FP8, W4A16, NVFP4, and GGUF variants.
Permissive model license
OpenMDW-1.1 allows use, modification, redistribution, and commercialization subject to its notice and termination terms.
Process
How the Trinity-Large-Thinking workflow works
Step 1
Select hosted or self-managed access
Prototype on Arcee Chat or OpenRouter before committing to the infrastructure required for self-hosting.
Step 2
Define narrow tools
Give each tool a clear schema, least-privilege permissions, deterministic validation, and an explicit failure path.
Step 3
Enable reasoning correctly
Configure the provider or vLLM parser so reasoning and final content are returned in the expected fields.
Step 4
Preserve the full assistant turn
Append reasoning_content, content, tool calls, and tool results to message history for every agent step.
Step 5
Control context growth
Budget for thinking tokens and remove complete older turns when needed instead of stripping recent reasoning.
Step 6
Evaluate on real trajectories
Measure tool-selection accuracy, malformed calls, recovery behavior, task completion, latency, and total token cost.
Step 7
Gate external actions
Require confirmation, policy checks, idempotency, and audit logs before tools create irreversible or high-impact changes.
Cost
Trinity-Large-Thinking pricing and free plan
The weights are downloadable without a model usage fee under OpenMDW-1.1, while infrastructure is separate. OpenRouter pricing varies by provider and routing; its current model page lists routed pricing around $0.22 per million input tokens, $0.85 per million output tokens, and $0.06 per million cached-read tokens.
OpenRouter routed API
$0.22 input / $0.85 output per 1M tokens
OpenRouter can route requests across available providers for price, speed, or reliability.
- Cached reads listed at $0.06 per 1M tokens
- Provider selection can change the exact price
- Reasoning tokens contribute to output usage
Arcee AI provider on OpenRouter
$0.25 input / $0.80 output per 1M tokens
Pin the request to Arcee AI's current hosted provider rate when consistency matters.
- Cached reads listed at $0.06 per 1M tokens
- Current provider availability can change
- OpenRouter account and billing required
Downloadable weights
No model usage fee
Run the official or quantized checkpoints on infrastructure you control.
- OpenMDW-1.1 terms apply
- Retain the license and applicable origin notices when redistributing
- Compute, storage, networking, and operations costs are separate
Arcee Chat
No public rate card
Arcee links to a browser chat for trying the model, but the model card does not publish a dedicated plan price.
- Useful for manual evaluation
- Check the live service for current limits
- Not a substitute for API cost testing
Pricing checked . Check current pricing at the source ↗
Assessment
Trinity-Large-Thinking strengths and limitations
Where it stands out
- Strong focus on multi-step agents and tool use rather than generic chat alone
- Sparse activation reduces work per token relative to a dense 398B model
- Open weights and a permissive license support commercial customization
- Long native context can retain substantial agent state
- Hosted API pricing is low relative to many closed frontier reasoning models
- Official quantized checkpoints widen deployment options
- Detailed integration guidance explains how to preserve reasoning across turns
What to consider
- The 398B total checkpoint is still expensive and complex to self-host
- OpenRouter currently exposes about half the model card's 512K context window
- Reasoning traces increase context growth, latency, storage, and output-token cost
- Dropping reasoning_content can degrade multi-step behavior and produce malformed tool calls
- Explicit reasoning may expose sensitive intermediate data if applications log or display it carelessly
- Arcee's published benchmarks are point-in-time tests and do not guarantee performance in a custom agent harness
- The model trails some closed and open competitors on several general reasoning and coding benchmarks in Arcee's own table
- Tool calls require application-level permission checks, validation, retries, and auditability
Compare
Trinity-Large-Thinking alternatives
The right alternative depends on the specific output, workflow, controls and budget your project requires.
Consumer
GLM-5.1
Evaluate GLM-5.1 for another open model focused on long-horizon agentic coding.
Explore GLM-5.1 →Consumer
Minimax M2.7
Consider MiniMax M2.7 when strong coding and agent benchmarks matter and a different hosted or open model stack is acceptable.
Explore Minimax M2.7 →Agents
Kimi K2.5
Use Kimi K2.5 for an alternative open agent-focused model with its own tool-use and coding strengths.
Explore Kimi K2.5 →Consumer
Claude Opus 4.6
Choose Claude Opus 4.6 when managed frontier performance and a larger commercial ecosystem matter more than open weights.
Explore Claude Opus 4.6 →Questions
Trinity-Large-Thinking FAQs
What is Trinity-Large-Thinking?
It is Arcee AI's reasoning-optimized 398B sparse mixture-of-experts model for planning, coding, tool calling, and multi-step agent workflows.
How many parameters does Trinity Large Thinking use?
The model has roughly 398B total parameters and activates about 13B per token through 4-of-256 expert routing plus a shared expert.
What is its context window?
Arcee's model card lists a 512K extended context window. The current OpenRouter endpoint lists 262,144 tokens, so applications must use the provider's actual limit.
How much does the API cost?
OpenRouter currently lists routed pricing around $0.22 per million input tokens, $0.85 per million output tokens, and $0.06 per million cached-read tokens. Exact provider rates can differ.
Can I use Trinity Large Thinking commercially?
Yes. OpenMDW-1.1 is permissive and allows commercial use, subject to retaining the license and applicable origin notices on redistribution and its other terms.
Why must I preserve reasoning_content?
Arcee says the model relies on its recent thinking to continue multi-step work. Omitting it can degrade agent behavior and cause malformed tool calls.
Can I self-host it?
Yes. Arcee publishes full and quantized checkpoints with vLLM, SGLang, Transformers, and Docker guidance, but the model's total size requires serious infrastructure planning.
Is Trinity Large Thinking a multimodal model?
The official model card describes a text-generation model. Some API gateways may accept additional content formats, but do not assume native image or audio understanding without endpoint-specific documentation and tests.
Bottom line
Our Trinity-Large-Thinking verdict
Trinity-Large-Thinking is a credible open-weight option for teams building tool-heavy agents and willing to implement reasoning preservation correctly. Its low hosted price, sparse activation, permissive license, long context, and official quantizations are attractive. The main risks are operational: a huge checkpoint, provider-specific context limits, fast-growing reasoning history, and the need to secure every tool action. Benchmark it inside the real harness rather than choosing from leaderboard scores alone.
Visit Trinity-Large-Thinking website ↗Similar Consumer Tools
View all Consumer AI tools →
Qwen3.5-Omni
Qwen 3.5 Omni - Alibaba's native omnimodal model with text, image, audio, and video understanding across 113 languages
LFM2.5-350M
LFM2.5-350M - Liquid AI's 350M-parameter edge model built for tool use and on-device agents
Cohere Transcribe
Cohere Transcribe- SOTA, free open-source speech recognition model
Gemma 4
Gemma 4 - Google's new open-weight reasoning and agentic model family across four sizes
AI training for the future of work.
Get access to all our AI courses, hundreds of real-world AI use cases, live expert-led workshops, an exclusive network of AI early adopters, and more.
AI Courses
Get unlimited access to all of our current & upcoming industry-specific AI courses for the duration of your subscription.
Daily Guides
To keep up with the rapid pace of AI, our team publishes AI implementation guides daily. Our library contains 300+ practical use cases to automate real-world work.
Workshops
Join weekly, live, interactive sessions with industry leaders who are at the forefront of AI for hands-on implementation guidance and exclusive insights.
Community
Network with an exclusive community of AI-first professionals who are working smarter with AI. Learn how early adopters are using AI in their work and businesses.